import librosa
import matplotlib.pyplot as plt 
import sys
from CQCC.cqcc import cqcc
import scipy.io.wavfile as wav
import soundfile as sf
import os
import numpy as np
import pickle
import argparse
import pandas as pd
import re
import matplotlib.pyplot as plt
import seaborn as sns
import random

CQCC in Python

filename = "./D18_1000001.wav"

Input Signal

in some python file they use librosa to read sound files

In my feature extraction code, they used soundfile to read, so i used that here for fair comparison, but both give same result

x,fs = sf.read(filename)
x = x.reshape(x.shape[0], 1)  # for one-channel signal 

Parameters

Same parameters are used in matlab code

B = 96
fmax = fs/2
fmin = fmax/2**9
d = 16
cf = 19
ZsdD = 'ZsdD'

Compute CQCC Features

The cqcc function return 7 results, I used 'CQcc' for Model building, so did the Baseline paper, and the python converted code in model building paper.

CQcc, LogP_absCQT, TimeVec, FreqVec, Ures_LogP_absCQT, Ures_FreqVec, absCQT = cqcc(x, fs, B, fmax, fmin, d, cf, ZsdD)
print("cqcc_feat:", CQcc.shape)       # number of frames * number of cep
print("cqcc_lpms:", LogP_absCQT.shape) 

cqcc_feat: (470, 60)
cqcc_lpms: (470, 863)

feats_py = [CQcc, LogP_absCQT, TimeVec, FreqVec, Ures_LogP_absCQT, absCQT]

CQCC in Matlab

I ran the cqcc code in matlab and saved each result from cqcc function in a text file.

'Ures_FreqVec_test.txt' -> I removed this result, because python version returns nothing in this variable.

matlab_output_folder = '/Users/asimadnan/Desktop/Mres/ASVspoof-main/CQCC/D18_1000001/'
features = ['CQcc_test.txt', 
           'LogP_absCQT_test.txt', 
           'TimeVec_test.txt', 
           'FreqVec_test.txt', 
           'Ures_LogP_absCQT_test.txt', 
           'absCQT_test.txt']

Regex to clean data, some rows had '\n' in them

RE = '[^0-9\.\-e]'

Helper functions

extract_from_matlab: This function reads txt files generated by matlab and saves them in ndarrays. plt_heatmap: to plot heatmap of a single feature

def extract_from_matlab(file_path):
    data = []
    with open(file_path) as input_file:
      for row in input_file:
        data.append( [float(re.sub(RE, '', i)) for i in row.split(',')] )
    data=np.array([np.array(xi) for xi in data])
    return data

# function to plot single heatmap of a feature.
def plt_heatmap(data,name):
    ax = sns.heatmap(data)
    plt.title('HeatMap of ' + name)
    plt.show()
    

NOTE: 2 results returned from matlab are transposed, when comparing there shape with python features

transposing them so the shape matches the features from python

feats_mat = []
for feature in features:
    if(feature == 'LogP_absCQT_test.txt' or feature == 'CQcc_test.txt'):
        feats_mat.append( extract_from_matlab( matlab_output_folder + feature).T) 
    else:
        feats_mat.append( extract_from_matlab( matlab_output_folder + feature)) 

Comparing result shapes

print('MATLAB            PYTHON     Match')
for idx,feature in enumerate(feats_mat):
    #print(feature.shape,feats_py[idx].shape)
    match = (feature.shape == feats_py[idx].shape)
    print('{mat:>12}    {pyt:>12} {match}'.format(mat=str(feature.shape), pyt=str(feats_py[idx].shape), match=match))
    

MATLAB            PYTHON     Match
   (470, 60)       (470, 60) True
  (470, 863)      (470, 863) True
    (1, 470)        (1, 470) True
    (1, 863)        (1, 863) True
 (8059, 470)     (8059, 470) True
  (863, 470)      (863, 470) True

heatmaps of all features of a sample file

# plotting all features side by side for comparison
for idx,feature in enumerate(feats_mat):
    plt.figure(figsize=(20, 20)) 

    plt.subplot(6, 2, idx + (idx +1))
    ax = sns.heatmap(feature)
    plt.title('MATLAB - HeatMap of ' + features[idx])
    
    #plt.figure(figsize=(12, 5)) 
    plt.subplot(6, 2, idx + (idx +2))
    ax = sns.heatmap(feats_py[idx])
    plt.title('PYTHON - HeatMap of ' + features[idx])
   
    

creating a function that plots all features of file side by side for comparison

def compare_features(filename,audio_file_location):
    #Python feature extraction
    #audio_file_location = './'
    x,fs = librosa.load(audio_file_location + filename, sr = 16000); 
    x = x.reshape(x.shape[0], 1) 
    B = 96
    fmax = fs/2
    fmin = fmax/2**9
    d = 16
    cf = 19
    ZsdD = 'ZsdD'
    CQcc, LogP_absCQT, TimeVec, FreqVec, Ures_LogP_absCQT, Ures_FreqVec, absCQT = cqcc(x, fs, B, fmax, fmin, d, cf, ZsdD)
    feats_py = [CQcc, LogP_absCQT, TimeVec, FreqVec, Ures_LogP_absCQT, absCQT]
    
    #Matlab features
    # a folder with name = filename will have  txt files for each feature 
    matlab_feature_folder = '/Users/asimadnan/Desktop/Mres/ASVspoof-main/CQCC/' + filename.split('.')[0] + '/'
    features = ['CQcc_test.txt', 
           'LogP_absCQT_test.txt', 
           'TimeVec_test.txt', 
           'FreqVec_test.txt', 
           'Ures_LogP_absCQT_test.txt', 
           'absCQT_test.txt']
    feats_mat = []
    for feature in features:
        if(feature == 'LogP_absCQT_test.txt' or feature == 'CQcc_test.txt'):
            feats_mat.append( extract_from_matlab( matlab_feature_folder + feature).T) 
        else:
            feats_mat.append( extract_from_matlab( matlab_feature_folder + feature)) 
    
    print('MATLAB            PYTHON     Match')
    for idx,feature in enumerate(feats_mat):
        match = (feature.shape == feats_py[idx].shape)
        print('{mat:>12}    {pyt:>12} {match}'.format(mat=str(feature.shape), pyt=str(feats_py[idx].shape), match=match))

    
    for idx,feature in enumerate(feats_mat):
        plt.figure(figsize=(20, 20)) 

        plt.subplot(6, 2, idx + (idx +1))
        ax = sns.heatmap(feature)
        plt.title('MATLAB - HeatMap of ' + features[idx])

        #plt.figure(figsize=(12, 5)) 
        plt.subplot(6, 2, idx + (idx +2))
        ax = sns.heatmap(feats_py[idx])
        plt.title('PYTHON - HeatMap of ' + features[idx])
  


    

selecting random files from training data to compare

train_path = '/Users/asimadnan/Desktop/Mres/ASVSPOOF_DATA/LA/ASVspoof2019_LA_train/flac/'
all_training_files = os.listdir(train_path)

random.choices(all_training_files,k=5)

['LA_T_1203196.flac',
 'LA_T_8682959.flac',
 'LA_T_3316254.flac',
 'LA_T_1846330.flac',
 'LA_T_4000361.flac']

ran matlab code on these files and saved features in txt files

# saving radnom files names
# randomom_files = ['LA_T_1203196.flac',
#  'LA_T_8682959.flac',
#  'LA_T_3316254.flac',
#  'LA_T_1846330.flac',
#  'LA_T_4000361.flac']

compare_features(randomom_files[0],train_path)

MATLAB            PYTHON     Match
   (389, 60)       (389, 60) True
  (389, 863)      (389, 863) True
    (1, 389)        (1, 389) True
    (1, 863)        (1, 863) True
 (8059, 389)     (8059, 389) True
  (863, 389)      (863, 389) True

compare_features(randomom_files[1],train_path)

MATLAB            PYTHON     Match
   (329, 60)       (329, 60) True
  (329, 863)      (329, 863) True
    (1, 329)        (1, 329) True
    (1, 863)        (1, 863) True
 (8059, 329)     (8059, 329) True
  (863, 329)      (863, 329) True

compare_features(randomom_files[2],train_path)

MATLAB            PYTHON     Match
   (225, 60)       (225, 60) True
  (225, 863)      (225, 863) True
    (1, 225)        (1, 225) True
    (1, 863)        (1, 863) True
 (8059, 225)     (8059, 225) True
  (863, 225)      (863, 225) True

compare_features(randomom_files[3],train_path)

MATLAB            PYTHON     Match
   (414, 60)       (414, 60) True
  (414, 863)      (414, 863) True
    (1, 414)        (1, 414) True
    (1, 863)        (1, 863) True
 (8059, 414)     (8059, 414) True
  (863, 414)      (863, 414) True

compare_features(randomom_files[4],train_path)

MATLAB            PYTHON     Match
   (214, 60)       (214, 60) True
  (214, 863)      (214, 863) True
    (1, 214)        (1, 214) True
    (1, 863)        (1, 863) True
 (8059, 214)     (8059, 214) True
  (863, 214)      (863, 214) True

Visually, CQCC and Ures LogP abs CQT are very different, rest of them look very similar.

Checking difference between each feature, and plotting the heatmap.

for idx, feat in enumerate(features):
        plt_heatmap((feats_mat[idx] - feats_py[idx]),feat)

Checking if Feature extraction process is deterministic, Python

i.e if we extract feature from same file multiple times, do they change or are they always same.

def extract_twice_python(filepath):
    #Python feature extraction
    #audio_file_location = './'
    x,fs = librosa.load(filepath, sr = 16000); 
    x = x.reshape(x.shape[0], 1) 
    B = 96
    fmax = fs/2
    fmin = fmax/2**9
    d = 16
    cf = 19
    ZsdD = 'ZsdD'
    CQcc, LogP_absCQT, TimeVec, FreqVec, Ures_LogP_absCQT, Ures_FreqVec, absCQT = cqcc(x, fs, B, fmax, fmin, d, cf, ZsdD)
    feats_1_py = [CQcc, LogP_absCQT, TimeVec, FreqVec, Ures_LogP_absCQT, absCQT]
    
    x,fs = librosa.load(filepath, sr = 16000); 
    x = x.reshape(x.shape[0], 1) 
    B = 96
    fmax = fs/2
    fmin = fmax/2**9
    d = 16
    cf = 19
    ZsdD = 'ZsdD'
    CQcc, LogP_absCQT, TimeVec, FreqVec, Ures_LogP_absCQT, Ures_FreqVec, absCQT = cqcc(x, fs, B, fmax, fmin, d, cf, ZsdD)
    feats_2_py = [CQcc, LogP_absCQT, TimeVec, FreqVec, Ures_LogP_absCQT, absCQT]
    
    return feats_1_py,feats_2_py
    
    
    

py_1, py_2 = extract_twice_python(train_path + randomom_files[0])

for idx, feat in enumerate(features):
        plt_heatmap((py_1[idx] - py_2[idx]),feat)

for idx, feat in enumerate(features):
        print( np.count_nonzero(py_1[idx] - py_2[idx]) )

0
0
0
0
0
0

Checking if deterministic in MATLAB

def extract_twice_matlab(filename):
    matlab_feature_folder = '/Users/asimadnan/Desktop/Mres/ASVspoof-main/CQCC/' + filename.split('.')[0] + '/'
    features = ['CQcc_test.txt', 
           'LogP_absCQT_test.txt', 
           'TimeVec_test.txt', 
           'FreqVec_test.txt', 
           'Ures_LogP_absCQT_test.txt', 
           'absCQT_test.txt']
    feats_mat_1 = []
    for feature in features:
        if(feature == 'LogP_absCQT_test.txt' or feature == 'CQcc_test.txt'):
            feats_mat_1.append( extract_from_matlab( matlab_feature_folder + feature).T) 
        else:
            feats_mat_1.append( extract_from_matlab( matlab_feature_folder + feature)) 
    
    matlab_feature_folder = '/Users/asimadnan/Desktop/Mres/ASVspoof-main/CQCC/' + filename.split('.')[0] + '_dup/'
    features = ['CQcc_test.txt', 
           'LogP_absCQT_test.txt', 
           'TimeVec_test.txt', 
           'FreqVec_test.txt', 
           'Ures_LogP_absCQT_test.txt', 
           'absCQT_test.txt']
    feats_mat_2 = []
    for feature in features:
        if(feature == 'LogP_absCQT_test.txt' or feature == 'CQcc_test.txt'):
            feats_mat_2.append( extract_from_matlab( matlab_feature_folder + feature).T) 
        else:
            feats_mat_2.append( extract_from_matlab( matlab_feature_folder + feature)) 

    return feats_mat_1,feats_mat_2
    

filename = 'LA_T_1846330.flac'
mat_1, mat_2 = extract_twice_matlab(filename)

for idx, feat in enumerate(features):
        plt_heatmap((mat_1[idx] - mat_2[idx]),feat)

for idx, feat in enumerate(features):
        print( np.count_nonzero(py_1[idx] - py_2[idx]) )

0
0
0
0
0
0